Asst. Prof. of AI/CogSci @HEC_Montreal and @Mila_Quebec; Prev PhD @Stanford Studying language learning and understanding in humans and AIs! She/They
Montreal, Canada
- Tweets 110
- Following 201
- Followers 657
- Likes 326
Pinned Tweet
Here’s an open source link to our new Nature Computational Science paper: nature.com/articles/s43588-0…
There's been lots of discussion on generative linguistics and generative AI. @EvaPortelance and I chime in here, suggesting that Neural Language Models (NLMs) reinforce key tenets of Chomsky’s approach in at least 3 fundamental ways, and that in turn, generative linguistics can
4
15
2,335
Friends! I am looking for PhD and Masters students to join my group for Fall 2025, Please share with students you may know who are interested in the intersection of language, CogSci, and AI! first deadline is Dec 1st, see my page for more info: evaportelance.github.io/pros…
7
146
402
56,453
Ever wondered how modern neural networks like LLMs can be useful to language acquisition, and more broadly cognitive science, if they are not a priori designed to be cognitive models? Well, @linguistMasoud and I answer this question in our new paper! osf.io/preprints/psyarxiv/b6…
1
13
82
23,062
Join me and other incredible speakers, @kmahowald @rljfutrell @cocosci_lab @glupyan @cdyangupenn @tyrell_turing @Brown_NLP to name a few, at the very cool (and free!) LLMs: Science & Stakes virtual summer school June 3-14 2024 to talk AI and CogSci!
skywritingspress.ca/#Speaker…
2
16
78
11,494
BIG NEWS: I will be joining the department of decision sciences of @HEC_Montreal as an Assistant Professor of machine learning this summer!! I'm excited to bring cognitive science and ML together at one of the best business schools around, and in my all time favorite city!
11
5
60
5,208
🎉New paper our in Cognitive Science @cogsci_soc with @mcxfrank and @jurafsky !
It's a deepdive into how visually-grounded LMs - and by extension possibly humans - learn the meanings of complex words like and/or, behind/in front, more/fewer.
(1/7)
2
15
56
9,379
New preprint by @linguistMasoud and I 🎉 Connecting seminal ideas from generative linguistics (theory adequacy, procedures, learnability and UG) to LMs and grammar induction models - discussing how AI and theoretical linguistics may serve each other: arxiv.org/abs/2411.10533
1
10
52
4,303
New paper out 🎉 !! @linguistMasoud and I wrote a perspective on how modern LMs and AI models can be useful for studying language acquisition in kids. Check out the open source paper in Language and Linguistics Compass: dx.doi.org/10.1111/lnc3.7000… and this original thread:
Ever wondered how modern neural networks like LLMs can be useful to language acquisition, and more broadly cognitive science, if they are not a priori designed to be cognitive models? Well, @linguistMasoud and I answer this question in our new paper! osf.io/preprints/psyarxiv/b6…
1
9
48
6,524
Our new open-access paper with Yuguang Duan, @mcxfrank, and @glupyan is out in @cogsci_soc 's Cognitive Science journal entitled "Predicting age of acquisition for children's early vocabulary in five languages using language model surprisal"
onlinelibrary.wiley.com/doi/…
(1/6)
1
12
46
6,334
As someone who grew up in a bilingual community and learnt 2 first langs, I was fascinated to discover that I was terrible at translating back and forth when asked, no connections existing between the lexicons. “What is ‘truck’ in French?” 🤷 could take me 10s to process.
34
1,886
Exciting stuff: I have a new paper out which was recently accepted at #CoNLL2021 called 'The Emergence of the Shape Bias results from Communicative Efficiency'. arxiv.org/abs/2109.06232 This is work with @jurafsky @mcxfrank @murefil and @LarocheRomain.
5
11
32
Our new paper is out! Why crediting the right reasoning steps is so important in RL LLM finetuning for difficult reasoning tasks. Congrats @a_kazemnejad @MAghajohari !
VinePPO, a straightforward modification to PPO, unlocks RL’s true potential for LLM Reasoning.
It beats RL-free methods (DPO and RestEM) and PPO, surpassing it in less steps(up to 9x), less time(up to 3x), and less KL with half memory.
Time to rethink RL post-training🧵: [1/n]
1
2
26
4,316
I am quoted in the @nytimes! Great article on the BabyLM Challenge and where the possible links between language models and human cognition currently stand.
nytimes.com/2023/05/30/scien…
#ai #languagemodels #languagelearning #cognition
4
18
2,051
Multimodal reasoning is needed for better image editing models and beyond: check out @benno_krojer new dataset and model that try to tackle this very issue!
Did you miss the recent Auroras? No problem! ✨🎆
Super excited to share AURORA, a *general* image editing model + high-quality data that improves where prev work fails the most:
Performing *action or movement* edits, i.e. a kind of world model setup
Insights/Details ⬇️
1
10
881
Finally, we survey recent work by up and coming researcher having adopted each of our modelling approaches and address the importance of computational modelling in language acquisition studies. @Adinseg @mitjanikolaus @xenia_ohmer @JLiu_Compuling @wkvong among other cool people!
7
343
The first step to joining my group is submitting a supervision request at Mila - Quebec AI Institute ! Applications now open 🎉 For next steps see evaportelance.github.io/pros…
Mila's annual supervision request process opens on October 15 to receive MSc and PhD applications for Fall 2025 admission! Join our community! More information here mila.quebec/en/prospective-s…
6
397
We offer proof-of-concept evidence that nuanced interpretations of function words are learnable from visually grounded contexts and non-symbolic general statistical learning, without prior knowledge/bias.
Check out the paper, it's open-source! doi.org/10.1111/cogs.13448
(7/7)
5
272
A little late with this one: I was quoted with others, including @LakeBrenden, in a @sciam article on AI and the prospects of developing better smaller models by @lauren_leffer , check it out! scientificamerican.com/artic…
5
261
Great team in a great city looking to hire!
The ML team at @MSFTResearch Montréal 🍁 is hiring a Senior Researcher with a background in ML / NLP!!!
Come work with us at the intersection of interactivity, modularity and reasoning in foundation models 😊
MSR is a highly collaborative environment where risky ideas are cherished.
Visit aka.ms/AAta22o, and please help us share this post in your networks!
Thread below 👇:
4
416
If your open to philosophical discussions (which I think are important even for CS students!) about LLMs vs human reasoning and language learning and their limitations, I would suggest these two: sciencedirect.com/science/ar…
and
arxiv.org/pdf/2106.08694
3
144
As developments towards natural language understanding and generation have improved leaps and bounds, with models like GPT-4, many have started asking how they can inform our understanding of human language acquisition.
1
2
364
1. How do VQA models learn to function words and do their representations generalize to unseen linguistic and visual contexts?
Using a set of probe tasks, we show that they learn gradient semantics for function words requiring spatial and numerical reasoning, as people do.
(3/7)
1
2
201
It is critical to examine how in practice linking hypotheses between models and human learners can be safely established. Here, we propose 4 modeling approaches, each having differing goals, from exploratory hypothesis generation to hypothesis differentiation and testing.
1
2
275
3. Do models learn these words in a similar order to kids; what drives these ordering effects?
Kids learn words in a consistent order, why? We find that learning difficulty is mainly driven by frequency of exposure, consistent with usage-based theories of acquisition.
(6/7)
1
2
326
On such approach is using LLM type models as `proofs of concept' for the learnability of linguistic knowledge (@a_stadt, @sleepinyourhat , Lisa Pearl, Marco Baroni, and others), but as we discuss in this paper, there are other approaches that are just as important!
1
2
234
Core to our argument, we show how the goals of these approaches align with the overarching goals of science and linguistics by connecting our taxonomy to the realist vs. instrumentalist approaches in philosophy of science.
1
2
320
Interpreting function words requires logical, numerical, and relational reasoning. Prior acquisition theories often relied on positing some innate knowledge/bias. Yet, VQA models can seemingly interpret these words without prior learning biases. We answer 3 research Qs:
(2/7)
1
1
211
Here, we investigate whether predictability relates to when children start producing different words (Age of Acquisition; AoA). We operationalized predictability in terms of a word’s surprisal in child directed speech, computed using n-gram and LSTM language models.
(3/6)
1
1
58
Second, we show that pressure brought on by communicative need is also necessary for it to persist across generations; simply having a shape bias in an agent's input language is insufficient.
1
2.Does the existence of alternative expressions affect their acquisition or are the meanings of function words acquired in isolation?
We find that models are sensitive to alternative expressions when interpreting language,
(4/7)
1
1
158
What makes a word easy to learn? Early-learned words are frequent and tend to name concrete referents. But words typically do not occur in isolation. Some words are predictable from their contexts; others less so.
(2/6)
1
1
70
There will be panel discussions throughout the two weeks and online participants who share their name and affiliation I believe will be able to ask questions both during talks and panels. Beyond that I don't know if any social event planned as of yet.
1
62
Eg. exclusive interpretations of or increase with the presence of the alternative and, early evidence supporting the acquisition of a fundamental skill for eventual pragmatic reasoning.
(5/7)
1
1
176